OBSERVATIONAL CONSTRAINTS ON THE NATURE OF DARK ENERGY: FIRST COSMOLOGICAL RESULTS FROM THE ESSENCE SUPERNOVA SURVEY

Transcription

1 The Astrophysical Journal, 666:694Y715, 2007 September 10 # The American Astronomical Society. All rights reserved. Printed in U.S.A. A OBSERVATIONAL CONSTRAINTS ON THE NATURE OF DARK ENERGY: FIRST COSMOLOGICAL RESULTS FROM THE ESSENCE SUPERNOVA SURVEY W. M. Wood-Vasey, 1 G. Miknaitis, 2 C. W. Stubbs, 1,3 S. Jha, 4,5 A. G. Riess, 6,7 P. M. Garnavich, 8 R. P. Kirshner, 1 C. Aguilera, 9 A. C. Becker, 10 J. W. Blackman, 11 S. Blondin, 1 P. Challis, 1 A. Clocchiatti, 12 A. Conley, 13 R. Covarrubias, 10 T. M. Davis, 14 A. V. Filippenko, 4 R. J. Foley, 4 A. Garg, 1,3 M. Hicken, 1,3 K. Krisciunas, 8,15 B. Leibundgut, 16 W. Li, 4 T. Matheson, 17 A. Miceli, 10 G. Narayan, 1,3 G. Pignata, 12 J. L. Prieto, 18 A. Rest, 9 M. E. Salvo, 11 B. P. Schmidt, 11 R. C. Smith, 9 J. Sollerman, 14,19 J. Spyromilio, 16 J. L. Tonry, 20 N. B. Suntzeff, 9,15 and A. Zenteno 9 Received 2006 November 21; accepted 2007 April 2 ABSTRACT We present constraints on the dark energy equation-of-state parameter, w ¼ P/(c 2 ),using60sneiafromthe ESSENCE supernova survey. We derive a set of constraints on the nature of the dark energy assuming a flat universe. By including constraints on ( M, w) from baryon acoustic oscillations, we obtain a value for a static equation-of-state parameter w ¼ 1:05 þ0:13 0:12 (stat 1 ) 0:13 (sys) and M ¼ 0:274 0:020 þ0:033 (stat 1 ) with a bestfit 2 /dof of These results are consistent with those reported by the Supernova Legacy Survey from the first year of a similar program measuring supernova distances and redshifts. We evaluate sources of systematic error that afflict supernova observations and present Monte Carlo simulations that explore these effects. Currently, the largest systematic with the potential to affect our measurements is the treatment of extinction due to dust in the supernova host galaxies. Combining our set of ESSENCE SNe Ia with the first-results Supernova Legacy Survey SNe Ia, we obtain a joint constraint of w ¼ 1:07 0:09 þ0:09 (stat 1 ) 0:13 (sys), M ¼ 0:267 0:018 þ0:028 (stat 1 ) with abest-fit 2 /dof of The current global SN Ia data alone rule out empty ( M ¼ 0), matter-only M ¼ 0:3, and M ¼ 1universesat>4.5. The current SN Ia data are fully consistent with a cosmological constant. Subject headings: cosmological parameters cosmology: observations supernovae: general Online material: color figures 1. INTRODUCTION: SUPERNOVAE AND COSMOLOGY We report on the analysis of 60 Type Ia supernovae (SNe Ia) discovered in the course of the ESSENCE program (Equation of State: SupErNovae trace Cosmic Expansion; an NOAO Survey Program) from 2002 through The aim of ESSENCE is to measure the history of cosmic expansion over the past 5 billion years with sufficient precision to distinguish whether the dark energy is different from a cosmological constant at the w ¼0:1 level. Here we present our first results and show that we are well on our way toward that goal. Our present data are fully consistent with a w ¼ 1, flat universe, and our uncertainty in w, the parameter that describes the cosmic equation 1 Harvard-Smithsonian Center for Astrophysics, Cambridge, MA 02138; wmwood-vasey@cfa.harvard.edu. 2 Fermilab, Batavia, IL Department of Physics, Harvard University, Cambridge, MA Department of Astronomy, University of California, Berkeley, CA Kavli Institute for Particle Astrophysics and Cosmology, Stanford Linear Accelerator Center, MS 29, Menlo Park, CA Space Telescope Science Institute, Baltimore, MD Johns Hopkins University, Baltimore, MD Department of Physics, University of Notre Dame, Notre Dame, IN Cerro Tololo Inter-American Observatory, National Optical Astronomy Observatory, Casilla 603, La Serena, Chile. 10 Department of Astronomy, University of Washington, Seattle, WA Research School of Astronomy and Astrophysics, Australian National University, Mount Stromlo and Siding Spring Observatories, Weston Creek PO 2611, Australia. 694 of state, analyzed in the way we outline here, will shrink below 0.1 for models of constant w as the ESSENCE program is completed. Other approaches to using the luminosity distances have been suggested to constrain possible cosmological models. We here provide the ESSENCE observations in a convenient form suitable for testing a variety of models. 21 As reported in a companion paper (Miknaitis et al. 2007), ESSENCE is based on a supernova search carried out with the 4 m Blanco Telescope at the Cerro Tololo Inter-American Observatory (CTIO) with the prime-focus MOSAIC II 64 Megapixel CCD camera. Our search produces densely sampled R-band and I-band light curves for supernovae in our fields. As described by Miknaitis et al. (2007), we optimized the search to provide the best constraints on w, given fixed observing time and the properties of both the MOSAIC II camera and the CTIO 4 m 12 Pontificia Universidad Católica de Chile, Departamento de Astronomía y Astrofísica, Casilla 306, Santiago 22, Chile. 13 Department of Astronomy and Astrophysics, University of Toronto, Toronto, ON M5S 3H4, Canada. 14 Dark Cosmology Centre, Niels Bohr Institute, University of Copenhagen, DK-2100 Copenhagen, Denmark. 15 Department of Physics, Texas A&M University, College Station, TX European Southern Observatory, D Garching, Germany. 17 National Optical Astronomy Observatory, Tucson, AZ Department of Astronomy, Ohio State University, Columbus, OH Department of Astronomy, Stockholm University, AlbaNova, Stockholm, Sweden. 20 Institute for Astronomy, University of Hawaii, Honolulu, HI See /essence/.

2 DARK ENERGY FROM ESSENCE SURVEY 695 telescope. Spectra from a variety of large telescopes, including Keck, VLT, Gemini, and Magellan, allow us to determine supernova types and redshifts. We have paid particular attention to the central problems of calibration and systematic errors that, on completion of the survey in 2008, will be more important to the final precision of our cosmological inferences than statistical sampling errors for about 200 objects. This first cosmological report from the ESSENCE survey derives some properties of dark energy from the sample presently in hand, which is still small enough that the statistics of the sample size make a noticeable contribution to the uncertainty in dark energy properties. But our goal is to set out the systematic uncertainties in a clear way so that these are exposed to view and so that we can concentrate our efforts where they will have the most significant effect. To infer luminosity distances to the ESSENCE supernovae over the redshift interval 0.15Y0.70, we employ the relations developed for SNe Ia at low redshift (Jha et al and references therein) among their light-curve shapes, colors, and intrinsic luminosities. The expansion history from z 0:7tothe present provides leverage to constrain the equation-of-state parameter for the dark energy as described below. In x 1 we sketch the context of the ESSENCE program. In x 2 we show from a set of simulated light curves that this particular implementation of light-curve analysis is consistent, with the same cosmology emerging from the analysis as was used to construct the samples, and that the statistical uncertainty we ascribe to the inference of the dark energy properties is also correctly measured. This modeling of our analysis chain gives us confidence that the analysis of the actual data set is reliable and its uncertainty is correctly estimated. Section 3 delineates the systematic errors we confront, estimates their present size, and indicates some areas where improvement can be achieved. Section 4 describes the sample and provides the estimates of dark energy properties using the ESSENCE sample. The conclusions of this work are given in x Context Supernovae have been central to cosmological measurements from the very beginning of observational cosmology. Shapley (1919) employed supernovae against the island universe hypothesis, arguing that objects such as SN 1885A in the Andromeda Nebula would have M ¼ 16 mag, which was out of the question. Edwin Hubble noted a mysterious class of exceptional novae which attain luminosities that are respectable fractions of the total luminosities of the systems in which they appear ( Hubble 1929). These extrabright novae were dubbed supernovae by Baade & Zwicky (1934). Minkowski (1941) divided them into two classes based on their spectra: Type I supernovae (SNe I) have no hydrogen lines, while Type II supernovae (SNe II) show H and other hydrogen lines. The high luminosity and observed homogeneity of the first handful of SN I light curves prompted Wilson (1939) to suggest that they be employed for fundamental cosmological measurements, starting with time dilation of their characteristic rise and fall to distinguish true cosmic expansion from tired light. After the SN Ib and SN Ic subclasses were separated from the SNe Ia (for a review see Filippenko 1997), this line of investigation has grown more fruitful as techniques of photometry have improved and as the redshift range over which supernovae have been well observed and confirmed to have standard light-curve shapes and luminosities has increased ( Rust 1974; Leibundgut et al. 1996; Riess et al. 1997, 2004; Goldhaber et al. 2001; Foley et al. 2005; Hook et al. 2005; Conley et al. 2006; Blondin et al. 2006). Within the uncertainties, the results agree with the predictions of cosmic expansion and provide a fundamental test that the underlying assumption of an expanding universe is correct. Evidence for the homogeneity of SNe Ia comes from their small scatter in the Hubble diagram. Kowal (1968) compiled data for the first well-populated Hubble diagram of SNe I. The 1 scatter about the Hubble law was 0.6 mag, but Kowal presciently speculated that distances to individual supernovae might eventually be known to 5%Y10% and suggested that [i]t may even be possible to determine the second-order term in the redshiftmagnitude relation when light curves become available for very distant supernovae. Precise distances to SNe Ia enable tests for the linearity of the Hubble law and provide evidence for local deviations from the local Hubble flow, attributed to density inhomogeneities in the local universe (Riess et al. 1995, 1997; Zehavi et al. 1998; Bonacic et al. 2000; Radburn-Smith et al. 2004; Jha et al. 2007). While SN Ia cosmology is not dependent on the value of H 0,itis sensitive to deviations from a homogeneous Hubble flow, and these regional velocity fields may limit our ability to estimate properties of dark energy, as emphasized by Hui & Greene (2006) and Cooray & Caldwell (2006). Whether the best strategy is to map the velocity inhomogeneities thoroughly or to skip over them by using a more distant low-redshift sample remains to be demonstrated. We have used a lower limit of redshift z > 0:015 in constructing our sample of SNe Ia. The utility of SNe Ia as distance indicators results from the demonstration that the intrinsic brightness of each SN Ia is closely connected to the shape of its light curve. As the sample of wellobserved SNe Ia grew, some distinctly bright and faint objects were found. For example, SN 1991T (Filippenko et al. 1992b; Phillips et al. 1992) and SN 1991bg (Filippenko et al. 1992a; Leibundgut et al. 1993) were of different luminosity, and their light curves were not the same either. The possible correlation of the shapes of supernova light curves with their luminosities had been explored by Pskovskii (1977). More homogeneous photometry from CCD detectors, more extreme examples from larger samples, and more reliable distance estimators enabled Phillips (1993) to establish the empirical relation between lightcurve shapes and supernova luminosities. The Calán-Tololo sample (Hamuy et al. 1996) and the CfA sample (Riess et al. 1999; Jha et al. 2006b) of SNe Ia have been used to improve the methods for using supernova light curves to measure supernova distances. Many variations on Phillips s idea have been developed, including m 15 (Phillips et al. 1999), MLCS (Riess et al. 1996; Jha et al. 2007), DM15 (Prieto et al. 2006), stretch (Goldhaber et al. 2001), CMAGIC (Wang et al. 2003), and SALT (Guy et al. 2005, 2007). These methods are capable of achieving the 10% precision for supernova distances that Kowal (1968) foresaw 40 years ago. In the ESSENCE analysis, we have used a version of the Jha et al. (2007) method called MLCS2k2. We have compared it with the results of the Spectral Adaptive Lightcurve Template (SALT; Guy et al. 2005) light-curve fitter used by the SNLS (Astier et al. 2006, hereafter A06). This comparison provides a test: if the two approaches do not agree when applied to the same data, they cannot both be correct. As shown in x 2, SALTand this version of MLCS2k2, with our preferred extinction prior, are in excellent accord when applied to the same data. While gratifying, this agreement does not prove that they are both correct. Moreover, as described in x 4, the cosmological results depend somewhat on the assumptions about SN host galaxy extinction that are employed. This has been an ongoing problem in supernova cosmology. The work of Lira (1995) demonstrated the empirical fact that although SNe Ia have a range of colors at

3 696 WOOD-VASEY ET AL. Vol. 666 maximum light, they appear to reach the same intrinsic color about 30Y90 days past maximum light, independent of lightcurve shape. Riess et al. (1996) used dereddened SN Ia data to show that intrinsic color differences exist near maximum light, with fainter SNe Ia appearing redder than brighter objects, and then used this information to construct an absorption-free Hubble diagram. Given a good set of observations in several bands, the reddening for individual supernovae can then be determined and the general relations between supernova luminosity and the light-curve shapes in many bands can be established (Hamuy et al. 1996; Riess et al. 1999; Phillips et al. 1999). The initial detections of cosmic acceleration employed either these individual absorption corrections ( Riess et al. 1998) or a full-sample statistical absorption correction ( Perlmutter et al. 1999). Finding the best approach to this problem, whether by shifting observations to the infrared, limiting the sample to low-extinction cases, or making other restrictive cuts on the data, is an important area for future work. Some ways to explore this issue are sketched in x 4. Kowal (1968) recognized that second-order terms in cosmic expansion might be measured with supernovae once the precision and redshift range grew sufficiently large. More direct approaches with the Hubble Space Telescope (HST ) were imagined by Colgate (1979) and Tammann (1979). Tammann anticipated that HST photometry of SNe Ia at z 0:5 would lead to a direct determination of cosmic deceleration and that the time dilation of SN Ia light curves would be a fundamental test of the expansion hypothesis. Early attempts at high-redshift supernova detection were undertaken by a Danish group in 1986Y1988 through the European Southern Observatory ( ESO) 1.5 m telescope at La Silla Observatory. Their cyclic CCD imaging of the search fields used image registration, convolution and subtraction, and realtime data analysis (Hansen et al. 1987). Alas, the rate of SNe Ia in their fields was lower than they had anticipated, and only one SN Ia, SN 1988U, was discovered and monitored in 2 years of effort (Hansen et al. 1987; Norgaard-Nielsen et al. 1989). More effective searches by the Lawrence Berkeley National Laboratory ( LBNL) group exploiting larger CCD detectors and sophisticated detection software showed that this approach could be made practical and be used to find significant numbers of highredshift SNe Ia (Perlmutter et al. 1995). By 1995, two groups, the LBNL-based Supernova Cosmology Project (SCP) and the High-Z Supernova Search Team ( HZT; Schmidt et al. 1998), were working in this field. The first SN Ia cosmology results using seven high-redshift SNe Ia (Perlmutter et al. 1997) found a universe consistent with M ¼ 1, but subsequent work by the SCP (Perlmutter et al. 1998) and by the HZT (Garnavich et al. 1998) revised this initial finding to favor a lower value of M. At the 1998 January meeting of the American Astronomical Society both teams reported that the SN Ia results favored a universe that would expand without limit, but at that time neither team claimed that the universe was accelerating. The subsequent publication of stronger results based on larger samples by the HZT (Riess et al. 1998) and by the SCP (Perlmutter et al. 1999) provided a surprise. The supernova data showed that SNe Ia at z 0:5 were about 0.2 mag dimmer than expected in an open universe and pointed firmly at an accelerating universe (for first-hand accounts, see Overbye 1999, p. 426; Riess 2000; Filippenko 2001; Kirshner 2002; Perlmutter 2003; reviews are given by, e.g., Leibundgut 2001; Filippenko 2004, 2005b). The supernova route to cosmological understanding continues to improve. One source of uncertainty has been the small sample of very well observed low-redshift supernovae ( Hamuy et al. 1996; Riess et al. 1999). The most recent contribution is the summary of CfA data obtained in 1997Y2001 (Jha et al. 2006b), but significantly enhanced samples from the CfA (Hicken et al. 2006), the Katzman Automatic Imaging Telescope (KAIT; Li et al. 2000; Filippenko et al. 2001; Filippenko 2005a), the Carnegie SN Program (Hamuy et al. 2006), the Nearby Supernova Factory (Wood-Vasey et al. 2004; Copin et al. 2006), and the Sloan Digital Sky Survey II Supernova Survey (SDSS-II; Frieman et al. 2004; Dilday et al. 2005) are forthcoming. As the low-z sample approaches 200 objects, the size of the sample will cease to be a source of statistical uncertainty for the determination of cosmological parameters. As described in x 3, systematic errors of calibration and K-correction will ultimately impose the limits to understanding dark energy s properties, and we are actively working to improve these areas (Stubbs & Tonry 2006). Some of the potential sources of systematic error in the high-z sample have been examined. The fundamental assumption is that distant SNe Ia can be analyzed using the methods developed for the low-z sample. Since nearby samples show that the SNe Ia in elliptical galaxies have a different distribution in luminosity than the SNe Ia in spirals (Hamuy et al. 2000; Howell 2001; Gallagher et al. 2005; Neill et al. 2006; Sullivan et al. 2006b), morphological classification of the distant sample may provide some useful clues to help improve the cosmological inferences ( Williams et al. 2003). For example, Sullivan et al. (2003) showed that restricting the SCP sample to SNe Ia in elliptical galaxies gave identical cosmological results to the complete sample, which is principally in spiral galaxies. The possibility of gray dust raised by Aguirre (1999a, 1999b) was examined by Riess et al. (2000) and Nobili et al. (2005) through infrared observations of high-z supernovae and was put to rest by the very high redshift observations of Riess et al. (2004, 2007). Improved methods for handling the vexing problems of absorption by dust have been developed by Knop et al. (2003) and Jha et al. (2007). These questions are described in more detail in x 3.3. The question of whether distant supernovae have spectra that are the same as those of nearby supernovae has been investigated by Coil et al. (2000), Lidman et al. (2005), Matheson et al. (2005), Hook et al. (2005), Howell et al. (2005), and Blondin et al. (2006). Foley et al. (2005) then confirmed that spectra of distant SNe Ia evolve over the lifetime of the SN Ia in the same way as those of nearby SNe Ia. In all cases, the evidence points toward nearby SNe Ia behaving in the same way as distant ones, bolstering confidence in the initial results. This observed consistency does not mean that the samples are identical, only that the variations between the nearby and distant samples are successfully accounted for by the methods currently in hand. We do not know whether this will continue to be the case as future investigations press for more stringent limits on cosmological parameters (Albrecht et al. 2006). The highest redshift SN Ia data (Riess et al. 2004, 2007) show the qualitative signature expected from a mixed dark energy/dark matter cosmology. Specifically, they show that cosmic deceleration due to dark matter preceded the current era of cosmic acceleration produced by dark energy. The sign of the observed effect on supernova apparent magnitudes reverses: SNe Ia at z 0:5 appear 0.2 mag dimmer than expected in a coasting cosmology, but the very distant supernovae whose light comes from z > 1 appear brighter than they would in that cosmology. By itself, this turnover is a very encouraging sign that supernova cosmology does not founder on gray dust or even on a simple evolution of supernova properties with cosmic epoch. As part of this analysis, Riess et al. (2004) constructed the gold sample of high-z and low-z supernovae whose observations met reasonable criteria for inclusion in an analysis of all of the published light curves and

4 No. 2, 2007 DARK ENERGY FROM ESSENCE SURVEY 697 spectra using a uniform method of deriving distances from the light curves. The analysis of the gold sample provided an estimate of the time derivative of the equation-of-state parameter, w, for dark energy. These observations are very important conceptually because the simplest fact about the cosmological constant as a candidate for dark energy is that it should be constant with redshift (i.e., w 0 ¼ dw/dz ¼ 0). The observations are consistent with a constant dark energy over the redshift range out to z 1:6. Other forms of dark energy might satisfy the observed constraints, but this observational test is one that the cosmological constant could have failed. The next definitive advance in our understanding of w came from the SNLS analysis of 71 SNe Ia, which constrained constant models of w to w ¼ 1:023 0:09 (stat) 0:054 (sys) and was consistent with a flat universe dominated by a cosmological constant (A06). In the analysis of the ESSENCE data presented in x 4, we use the supernova data to constrain the properties of w, asfirstcarried out by White (1998) and Garnavich et al. (1998). This parameterization of dark energy by w is not the only possible approach. A more detailed approach is to compare the observational data to a specific model and, for example, try to reconstruct the dark energy scalar field potential (see, e.g., Li et al. 2007). A more agnostic view is that we are simply measuring the expansion history of the universe, and a kinematic description of that history in terms of expansion rate, acceleration, and jerk (Riess et al. 2004, 2007; Rapetti et al. 2007) covers the facts without assuming anything about the nature of dark energy. The ESSENCE project was conceived to tighten the constraints on dark energy at z 0:5 to reveal any discrepancy between the observations and the leading candidate for dark energy, the cosmological constant. A simple way to express this is that we aim for a 10% uncertainty in the value of w. This program is similar to the approach of the SNLS being carried out at the Canada-France-Hawaii Telescope, and we compare our methods and results to theirs (Guy et al. 2005; A06) at several points in the analysis below. The SNLS has taken the admirable step of publishing their light curves online and making the code of their light-curve fitting program, SALT, available for public inspection and use. 22 Making the light curves public, as was done for the results of the HZT and its successors (Riess et al. 1998; Tonry et al. 2003; Barris et al. 2004; Krisciunas et al. 2005; Clocchiatti et al. 2006), by Knop et al. (2003), by Riess et al. (2004, 2007) for the very high redshift HSTsupernova program, and for the low-z data of Hamuy et al. (1996), Riess et al. (1999), and Jha et al. (2006b), provides the opportunity for others to perform their own analysis of the results. In addition to exploring a variety of approaches to analyzing our own SN Ia observations, in x 4 we show the first joint constraints from ESSENCE and the first year of SNLS, as well as some joint constraints derived from combining these with the Riess et al. (2004) gold sample. 2. LUMINOSITY DISTANCE DETERMINATION The physical quantities of interest in our cosmological measurements are the redshifts and distances to a set of spacetime points in the universe. The redshifts come from spectra and the luminosity distances, D L, come from the observed flux of the supernova combined with our understanding of SN Ia light curves from nearby objects. Extracting a luminosity distance to a supernova from observations of its light curve necessitates a number of assumptions. 22 See /conf/release/. We use the observations of nearby SNe Ia to establish the relations between color, light-curve shape in multiple bands, and peak luminosity. These nearby observations attain high signalto-noise ratios (S/Ns), and the nearby objects can be observed in more passbands (including infrared) than faint, distant objects. We assume that the resulting method of converting light curves to luminosity distances applies at all redshifts. The observed spectral uniformity of supernovae over a range of redshift (Coil et al. 2000; Lidman et al. 2005; Hook et al. 2005; Blondin et al. 2006) supports this approach. We assume that R V, the ratio of extinction in the V band to the color excess E(B V ), is independent of redshift. In x 3.3 we test the potential systematic effect of departures from this assumption. We adopt an astrophysically sensible prior distribution of host galaxy extinction properties, with a redshift dependence that is derived from the simulations we present below. Our approach is to conduct comprehensive simulations of the ESSENCE data and analysis. As described by Miknaitis et al. (2007), we use this same approach to explore our photometric performance. For the aspects of our analysis that are downstream of the light-curve generation, we generate sets of synthetic light curves and subject them to our analysis pipeline. In this way we can test the performance of our distance-fitting tools, and by exaggerating various systematic errors (zero-point offsets, etc.), we can assess the impact of these effects on our determination of w Extracting Luminosity Distances from Light Curves: Distance Fitters We use the MLCS2k2 method of Jha et al. (2007) as the primary tool to derive relative luminosity distances to our SNe Ia. For comparison, we also provide the results obtained using the SALT fitter of Guy et al. (2005) on the ESSENCE light curves. SALT was used in the recent cosmological results paper from the SNLS (A06). We provide a consistent and comprehensive set of distances obtained to nearby, ESSENCE, and SNLS supernovae for each luminosity distance fitting technique. The ESSENCE light curves used in this analysis were presented by Miknaitis et al. (2007), and we provide them online, together with our set of previously published light curves for nearby SNe Ia, for the convenience of those interested (see footnote 21). Additional SN Ia light-curve fitting methods will be further explored in future ESSENCE analyses. Understanding the behavior of our distance determination method is critical to our goal of quantifying the uncertainties of our analysis chain. MLCS2k2andSALT,aswellasthelight-curve stretch approach used by Perlmutter et al. (1997, 1999), Goldhaber et al. (2001), and Knop et al. (2003), exploit the fact that the rate of decline, the color, and the intrinsic luminosity of SNe Ia are correlated. At present we treat SNe Ia as a single-parameter family, and the distance-fitting techniques use multicolor light curves to deduce a luminosity distance and host galaxy reddening for each supernova. Previous papers have shown that the different techniques produce relative luminosity distances that scatter by 0.10 mag for an individual SN Ia (e.g., Tonry et al. 2003), but this scatter is uncorrelated with redshift. Consequently, the cosmological results are insensitive to the distance fitting technique. However, as described by Miknaitis et al. (2007), the measurement of the equation-of-state parameter hinges on subtle distortions in the Hubble diagram, so we have undertaken a comprehensive set of simulations to understand potential biases introduced by MLCS2k2. Based on simulations of 10,000 light curves that explored the data quality range spanned by the nearby and ESSENCE light curves, we developed a set of quality thresholds

5 698 WOOD-VASEY ET AL. Vol. 666 TABLE 1 MLCS2k2 Fit Parameter Quality Cuts TABLE 2 SALT Fit Parameter Quality Cuts Fit Parameter Requirement Fit Parameter Requirement Degrees of freedom... dof :4 1:7 Time of maximum uncertainty... T maxerr 2:0 rest-frame days First observation with S/N > days Last observation with S/N > days Notes. See Table 4 for the MLCS2k2 fit parameters used for the cosmological analysis presented in this paper. These selection criteria were derived based on Monte Carlo simulations discussed in x 2.5. The number of degrees of freedom is the number of light-curve points with S/N > 5minusthefourindependent MLCS2k2 fit parameters: m V,, A V,andT max. to require for including SNe Ia in our cosmological analysis. These cuts are summarized in Table 1. The fit quality requirements derived from fitting these simulated light curves with the SALT fitter are summarized in Table 2. The MLCS2k2 approach (Riess et al. 1996, 1998; Jha et al. 2007) to determining luminosity distances uses well-observed nearby SNe Ia to establish a set of light-curve templates in multiple passbands. The parameters (roughly equivalent to the variation in peak visual luminosity, this parameter characterizes intrinsic color, rate of decline, and peak brightness), A V (the V-band extinction of the supernova light in its host galaxy), and (the distance modulus) are then determined by fitting each multiband set of distant supernova light curves to redshifted versions of these templates. Jha et al. (2007) present results from MLCS2k2 based on nearby SNe Ia. Here we have modified MLCS2k2 for application to both high- and low-redshift SNe Ia. We begin with a rest-frame model of the SN Ia in its host galaxy and then propagate the model light curves through the host galaxy extinction, K-correction, and Milky Way extinction to the detector, incorporating the measured passband response (including the atmosphere for ground-based observations). We then fit this model directly to the natural-system observations. This forward-modeling approach has particular advantages in application to the more sparsely sampled (in color and time) data typical of high-redshift SN searches. The SALT method of Guy et al. (2005), which was used for the SNLS first-results analysis of A06, constructs a fiducial SN Ia template using combined spectral and photometric information, then transforms this template into the rest frame of the SN Ia, and finally calculates a flux, stretch, and generalized color. The color parameter in SALT is notable in that it includes both the intrinsic variation in SN Ia color and the extinction from dust in the host galaxy within a single parameter (in contrast, MLCS2k2 attempts to separate these components of the observed colors for each supernova). While the reddening vector (attenuation vs. color excess) is similar to the SN Ia color versus absolute magnitude relation, the two sources of correlated color and luminosity variation are not identical. The stretch and color parameters of SALT were used by A06 to estimate luminosity distances by marginalizing over the stretchluminosity and color-luminosity relationships during the cosmological fit for the full SN Ia sample. Given that the SALT color parameter conflates the two physically distinct phenomena of host galaxy extinction and SN Ia color variation, it is remarkable and perhaps a source of deep insight that this treatment works as well as it does. Because of both survey selection effects and possible demographic shifts in the host environments of SNe Ia, we would not expect that the proportion of reddening from dust Degrees of freedom... dof 5 Stretch... 0:5 s < 1:4 Time of maximum uncertainty... T maxerr 2:0 rest-frame days Observations after B-band maximum... >1 First observation with S/N > days Notes. See Table 5 for the SALT fit parameters used for the cosmological analysis presented in this paper. These selection criteria were based on A06 with additional sanity checks on the stretch parameter and uncertainty in the time of maximum light. The number of degrees of freedom is the number of light-curve points with S/N > 5 minus the four independent SALT fit parameters: m B, stretch, color, and T max. and from intrinsic variation would remain constant with redshift as this approach assumes. However, the SALT/A06 method does seem to work quite well in practice Sensitivity to Assumptions about the Host Galaxy Extinction Distribution: Extinction Priors The best way to treat host galaxy extinction is a serious question for this work and for the field of supernova cosmology. The Bayesian approach we use is detailed in x 3.4. Here we describe simulations that are designed to evaluate the effects of those methods. There have been four basic approaches to combining reddening measurements with astrophysical knowledge to determine the host galaxy extinction along the line of sight: (1) assume that linear A V is the natural space for extinction and assume a flat prior (Perlmutter et al. 1999; Knop et al. 2003); (2) use models of the dust distribution in galaxies ( Hatano et al. 1998; Commins 2004; Riello & Patat 2005) to model line-of-sight extinction values (Riess et al. 1998, 2004; Tonry et al. 2003); (3) assume that the distribution of host galaxy A V follows an exponential form (Jha et al. 2007), based on observed distributions of A V in nearby SNe Ia; and (4) self-calibrate within a set of low-z SNe Ia to obtain a consistent color+a V relationship and assume that relation for the full set (A06). Approach 1 assumes the least prior knowledge about the distribution of A V and produces a Gaussian probability distribution for the fitted luminosity distance. However, this approach weakens the ability to separate intrinsic SN Ia color from A V, resulting in a fit parameter A V that is a mixture of the two. An A V that is truly related to the dust extinction should never be negative. The probability prior with 1 < A V < þ1 is not the natural range over which to assume a flat distribution; the physically reasonable prior on A V should be strictly positive. One approach is to base the prior for absorption on the distribution of dust in galaxies. Theoretical modeling of dust distributions in galaxies, such as that of Hatano et al. (1998), Commins (2004), and Riello & Patat (2005), provides a physically motivated dust distribution. This method represents approach 2 above and is the method we adopt here. In contrast, Jha et al. (2007) empirically derived an exponential A V distribution from MLCS2k2 fits to nearby SNe Ia by assuming a particular color distribution of SNe Ia. This distribution was derived using the empirical fact that SNe Ia reach a common color about 40 days past maximum light (Lira 1995). They found an exponential distribution of A V, p(a V ) / exp A V ; ð1þ

6 No. 2, 2007 DARK ENERGY FROM ESSENCE SURVEY 699 Fig. 1. Distribution of the MLCS2k2 light-curve width parameter and A V for the MLCS2k2 fits with the glosz prior to the nearby (dotted line), ESSENCE (solid line), and SNLS (dashed line) SNe Ia considered in this paper. The glosz prior (dot-dashed line) is shown here for z ¼ 0, where it is equivalent to the glos prior. Note that we are mixing two slightly different things in showing the prior with these estimated mean fit parameters. The prior, which directly relates to the mode, is not expected to match the a posteriori mean distribution of the fit parameters. See Fig. 7 for the ESSENCE selection effect as a function of redshift. See Table 4 for the full set of MLCS2k2 light-curve fit results for these SNe Ia. [See the electronic edition of the Journal for a color version of this figure.] where ¼ 0:46 mag. Unfortunately, the highest extinction objects drive the tail of this exponential and significantly affect the fit, resulting in a prior sensitive to sample selection, which differs significantly in high-redshift searches compared to the nearby objects studied by Jha et al. (2007). A06 analyzed the results of the SALT SN Ia light-curve fitter with approach 4 and have systematic sensitivities that are similar to those of approach 1. We use MLCS2k2 as our main analysis tool. We designate approach 1 the flatnegav prior and approach 3 the default prior and discuss both of these further in x 3.4. Approach 2 is based on a galactic line-of-sight ( glos ) prior on A V : ˆpðA V Þ / A exp A V þ 2B pffiffiffiffiffi exp A2 V ; ð2þ where A ¼ 1, B ¼ 0:5, ¼ 0:4, ¼ 0:1, and ˆp(A V ) 0for A V < 0. This exponential plus one-sided narrow Gaussian glos prior is based on the host galaxy dust models of Hatano et al. (1998), Commins (2004), and Riello & Patat (2005). As described below, we have modeled our selection effects with redshift to adapt the glos prior into the glosz prior that is the basis for our analysis. We feel that this approach leverages our best understanding of the effects of extinction and selection. Figures 1 and 2 show the distribution of the fit parameters and overlay the prior distribution assumed for each of these approaches. Figures 3 and 4 demonstrate that there is no significant residual correlation between distance modulus and the fit parameters using the built-in MLCS2k2 width-luminosity relationship or SALT using the and parameters from A06. Figure 5 compares the fit distances and extinction/color parameters of the MLCS2k2 glosz and SALT fit results for the ESSENCE, SNLS, and nearby samples. The distribution of recovered and A V match their imposed priors for MLCS2k2 glosz, while the stretch and color fit parameters from SALT show a consistent distribution for the three different sets of SNe Ia ESSENCE Selection Effects and the Motivation for a Redshift-dependent Extinction Prior We examined the effect of the survey selection function on the expected demographics of the ESSENCE SNe Ia and explored the interplay between extinction, Malmquist bias, and our observed light curves. To determine the impact of the selection bias, we developed a Monte Carlo simulation of the ESSENCE search. We created a range of supernova light curves that match the properties of the nearby sample, added noise based on statistics from actual ESSENCE photometry, and then fitted the resulting light curves in the same manner that the real events are analyzed. In this way we estimated the impact of subtle biases, although this simulation cannot test for errors in our lightcurve model or population drift with redshift. Based on its low-redshift training set, MLCS2k2 is able to output a finely sampled light curve given a redshift (z), distance modulus (), light-curve shape parameter (), host extinction (A V ), host extinction law (R V ), date of rest-frame B-band maximum light (t 0 ), Milky Way reddening [E(B V ) MW ], and the bandpasses of the observations. At a given redshift we calculated a distance modulus, true, from the luminosity distance for the standard cosmology ( m ¼ 0:3, ¼ 0:7), and that distance modulus plus an assumed M B ¼ 19:5magforSNeIasetthe brightness for our simulated supernovae. Varying the assumed cosmology does not significantly impact the simulation results since we are comparing the input distance modulus with the recovered distance modulus, obs, which is independent of the cosmology.

7 700 WOOD-VASEY ET AL. Vol. 666 Fig. 2. Distribution of the SALT light-curve stretch and the SALT estimated color plus extinction (the c parameter of Guy et al. [2005]; sixth column of Table 5) for the nearby (dotted line), ESSENCE (solid line), and SNLS (dashed line) SNe Ia considered in this paper. The priors for SALT are effectively flat for stretch and color, and SALT quotes minimum 2 values instead of the estimated mean parameter values of MLCS2k2. See Table 5 for the full set of SALT light-curve fit results for these SNe Ia. [See the electronic edition of the Journal for a color version of this figure.] At each of a series of fixed redshifts, we created 1000 simulated light curves with parameters chosen from random distributions. The light-curve width,, was selected from the Jha et al. (2007) distribution measured from the low-z sample. The distribution is approximately a Gaussian peaking at ¼ 0:15 with an extended tail out to ¼ 1:5. The host extinction for each simulated event, A V, was selected from either the Jha et al. (2007) distribution (default) estimated from the local sample or from a galaxy line-of-sight estimation (glos). The default distribution was an exponential decay with index 0.46magandsettozeroforA V < 0:0 mag. The glos distribution is also set to zero for A V < 0:0 mag and combines a narrow Gaussian with an exponential tail for A V > 0:0 mag (see eq. [2]). The extinction law is assumed to be R V ¼ 3:1. The Milky Way reddening [E(B V ) MW ] distribution was constructed from the Schlegel et al. (1998, hereafter SFD98) reddening maps that Fig. 3. Distance modulus () residuals with respect to a CDM cosmology as a function of the MLCS2k2 glosz fit parameters: and A V.SeeTable4.[See the electronic edition of the Journal for a color version of this figure.]

8 No. 2, 2007 DARK ENERGY FROM ESSENCE SURVEY 701 Fig. 4. Distance modulus () residuals with respect to a CDM cosmology as a function of the SALT fit parameters: stretch and color. See Table 5. [See the electronic edition of the Journal for a color version of this figure.] cover the ESSENCE fields. E(B V ) MW was measured for 10,000 random locations in each ESSENCE field, and the reddening was selected from the sum of the histograms (see Fig. 2). The dates of observation for a simulated SN Ia were based on the actual dates of ESSENCE 4 m observations. An ESSENCE field was chosen at random from the list of monitored fields and adateofmaximum,t 0, selected to fall randomly between the Modified Julian Date (MJD) of the first and last observation of an observing season. The simulated light curve was then interpolated for only those dates that ESSENCE took images. With each ESSENCE field observation, we estimated the magnitude in R and I that provided a 10 photometric detection based on the seeing and sky brightness. The S/N for each simulated lightcurve point was then scaled from the 10 detection magnitude, assuming that the noise was dominated by the sky background. For each date of ESSENCE observation, we have a simulated noiseless magnitude and an estimate of the S/N of the observation. To each simulated observation we added an appropriate random value in flux space selected from a normal distribution with a width corresponding to the predicted S/N. MLCS2k2 was then used to fit the simulated light curves and provide estimates of,, A V,andt 0,assumingafixedR V ¼ 3:1 in the same manner as used for the real light curves. MLCS2k2 required an initial guess of the date of maximum, an estimate achieved by selecting from a normal distribution about the true date with a 1 width of 2 days. The SFD98 Milky Way reddening was also required in MLCS2k2 and was provided from the true reddening after adding an uncertainty of 10%. Finally, in the real ESSENCE data we discarded supernovae when the MLCS2k2 reduced 2 /dof indicated a very poor fit. We treated the simulated light curves similarly and dropped events from the sample if the reduced 2 exceeded Deriving an Extinction Prior from the Simulation Results Simulated ESSENCE samples were created at a range of redshifts out to z ¼ 0:70, and the light curves that passed the detection criteria from the actual ESSENCE search were fitted with MLCS2k2. The fitting was done with the default prior and the glos prior (with corresponding A V distributions). The difference between the true (input) distance modulus and recovered (fit) distance modulus,, was calculated for each event, and the mean, median, and dispersion for the ensemble were calculated at each redshift. The median of the simulations was within 0.03 mag for z < 0:45, but at higher redshift the simulated supernovae were estimated to be brighter than the input supernovae by more than 0.2 mag. This bias results from the loss of faint events (large A V and large ) from the sample as the distance increases. In a sense, this is a classic Malmquist bias, but here it is caused by an uninformed prior. These results are shown in Figure 6. The decreasing ability to observe large A V events as the redshift increases (see Fig. 7) makes it clear that using a single A V prior for all redshifts is not correct. Because events with large A V and large are lost at high redshift due to the magnitude limits of the search, we should adjust the prior as a function of z to account for these predictable losses. Applying redshiftdependent window functions to the basic glos prior provides a much better prior as a function of redshift. We fit the recovered A V distributions derived from the simulations, which start with a uniform A V, to a window function based on the error function (integral of a Gaussian), and two parameters describe where that function drops to half its peak value (A 1/2 ) and the width of the transition ( A ). The window function W has the form W A V ; A 1=2 ; A 1 ¼ 1 p ffiffiffi Z (AV A 1=2 ) A 1 e x 2 dx; where A 1/2 and A are functions of z and estimated from the simulations. A similar process was applied to the distribution, and Table 3 provides the parameters. We embody this prescription in the glosz prior we use for our main MLCS2k2 light-curve fitting. The glosz prior is the glos prior modified by the above ð3þ

9 702 WOOD-VASEY ET AL. Vol. 666 Fig. 5. Distance modulus and A V as a function of redshift for MLCS2k2 glosz minus the SALT distance modulus and ; color ( ¼ 1:57) for the ESSENCE, SNLS, and nearby data sets. High z refers to SNe Ia with z 0:15, low z to z < 0:15. The dot-dashed line shows the weighted average of the difference for each quantity, while the dashed line shows the line of zero difference. While the luminosity distances are offset between the two fitters, this is mainly due to a slightly different definition of the M parameter that defines the absolute luminosity of an SN Ia and the Hubble constant. The relative average difference between low redshift and high redshift is mag. This agreement translates to a similar agreement in the cosmological parameters obtained with each approach (see Figs. 10 and 11). [See the electronic edition of the Journal for a color version of this figure.] window functions in A V and. The simulations using the glosz prior provide a median within 0.03 mag for all redshifts less than 0.7, which we judge to be satisfactory performance. The glosz prior corrects for selection biases that have a significant impact on recovering the true distance moduli at faint magnitudes and high redshift. The need for the glosz prior underscores the importance of understanding selection bias, dust extinction priors, and light-curve width distributions in determining distances at the few percent level in supernova surveys Comparison of MLCS2k2 and SALT Luminosity Distance Fitters The release of the source code to the SALT fitter (Guy et al. 2005) makes a modern SN Ia light-curve fitter fully accessible and available to the community. This public release of SALT allows us to compare the results of our MLCS2k2 distance fitter (presented in Table 4) with the SALT fitter used in the SNLS first-results paper (A06). We present the results of SALT fits to our nearby and ESSENCE samples in Table 5. To compute the distance moduli we quote in that table, we assume the ¼ 1:52, ¼ 1:57 values from A06. To calibrate the additional dispersion to add to the distance moduli of MLCS2k2 and SALT, we fitted a CDM model to the nearby sample alone and derived the additional add to add in quadrature to recover 2 /dof of 1 for the nearby sample. This add is related to the intrinsic dispersion of the absolute luminosity of SNe Ia but is not precisely the same both because the light-curve fitters include varying degrees of model uncertainty and because the light curves of the SNe Ia are subject to photometric uncertainty. We find add ¼ 0:10 mag for MLCS2k2 with the glosz prior and add ¼ 0:13 mag for SALT. These values should be added to the uncertainties given in Tables 4 and 5. Figure 5 visually demonstrates that the relative luminosity distances using the SALT light-curve fitter agree, within uncertainties, with the MLCS2k2 distances when the latter are fitted using the glosz A V prior Testing the Recovery of Cosmological Models Using Simulations of the ESSENCE Data Set In order to assess our ability to recover unbiased estimates of the cosmological parameters and the accuracy of our uncertainties, we have carried out a set of Monte Carlo simulations of our

10 No. 2, 2007 DARK ENERGY FROM ESSENCE SURVEY 703 TABLE 3 glosz Window Function Parameters z A 1/ 2 A 1/ Fig. 6. Median of the distance modulus error as a function of redshift for the simulated data sets. The points show the median value of the difference between the input true and recovered obs of about 1000 simulated supernovae at each redshift. The lines indicate the rms spread of the recovered distance modulus. data set. We generated 5000 instances of an SN Ia sample designed to be statistically similar to the actual ESSENCE plus nearby sample after application of the quality cuts in Table 1 (in terms of redshift distribution, peculiar velocities, distance modulus errors, etc.) around a fiducial cosmology (flat: M ¼ 0:3, ¼ 0:7).Wethenfittedeachofthesesamplestorecover ( M, w) in a flat universe in combination with the baryon Fig. 7. Recovered distribution of visual extinctions for simulated supernovae in the ESSENCE sample if the input distribution were uniform in A V out to large extinctions. The curves are fitted to determine the parameters of the window function (see Table 3), which is then used to modify the glos prior as a function of redshift into the glosz prior. We estimate the SNLS selection function as extending +0.2 in redshift deeper than the ESSENCE selection function. acoustic oscillation (BAO) constraint (Eisenstein et al. 2005). To handle the latter, we shifted the results from this paper to our fiducial cosmology and randomly varied its value around that point using provided uncertainties. We then compared the mean recovered cosmological parameters with the input values and found that they are biased at less than the 1% level. Furthermore, we find that the error estimates on the cosmological parameters are slightly overconservative at about the 0.5% level. Hence, our methodology is sufficiently accurate for our requirements. 3. POTENTIAL SOURCES OF SYSTEMATIC ERROR Here we identify and assess sources of systematic error that could afflict our measurements. These can be divided into two groups. Certain sources of systematic error may introduce perturbations either to individual photometric data points or to the distances or redshifts estimated to the SNe Ia. Others affect the data in a more or less random fashion and produce excess scatter in the Hubble diagram. Errors that are uncorrelated with either distance or redshift will not bias the cosmological result. These sources of photometric error are detailed by Miknaitis et al. (2007); we summarize those results here in Table 6. We add these effects in quadrature to the statistical uncertainties given by the luminosity distance fitting codes for each SN Ia phot scatter distance measurement: ¼ 0:026 mag. In x 2 we discussed our testing of the MLCS2k2 fitter on simulated data sets that replicate the data quality of the ESSENCE and nearby SNe Ia. We explore the issue of host galaxy extinction further in xx 3.3 and 3.4. The interaction of Malmquist bias and selection effects with the extinction and color distribution of SNe Ia is discussed in x 3.5. Any noncosmological difference in measurements of nearby and distant SNe Ia has the potential to perturb our measurement of w. Table 7 lists potential systematic effects of this sort. We present both our estimate of the sensitivity (dw/dx) ofthe equation-of-state parameter to each potential systematic effect and our best estimate of the potential size of the perturbation, x. The upper bound on the bias introduced in w is then w ¼ (dw/dx) x. Miknaitis et al. (2007) discuss the systematic uncertainties on, which we convert here to systematic uncertainties on w, due to photometric errors from astrometric uncertainty on faint objects (w ¼ 0:005), potential biases from the difference imaging (w ¼ 0:001), and linearity of the MOSAIC II CCD (w ¼ 0:005). None of these contributed noticeably to the systematic uncertainty in our measurement of w. The rest of this section describes how we appraised our additional potential sources of systematic uncertainty. The conclusion of this section is that our current overall estimate for the 1 equivalent systematic uncertainty in a